Granite 4.1: IBM's Advanced Open Source Language Model

Granite 4.1 is a family of open source language models developed by IBM, specifically designed for enterprise use. This model family includes three sizes: 3B, 8B, and 30B parameters, all licensed under Apache 2.0 and trained on an extensive dataset of 15 trillion tokens.

Key features include:

• Dense architecture - Unlike previous models, Granite 4.1 does not utilize mixture of experts (MoE) or extended reasoning chains, resulting in a more efficient processing design.
• Comprehensive training phases - The model underwent five distinct training phases, each with specific data mixtures and learning goals to enhance performance.
• Robust data quality pipeline - A filtering system was implemented to ensure high-quality training data, rejecting poor examples before fine-tuning.
• Reinforcement learning - Four rounds of reinforcement learning were conducted to improve instruction following and overall model performance.

The benchmarks demonstrate that Granite 4.1's 8B model consistently outperforms its predecessor, Granite 4.0-H-Small, across various tasks, indicating significant advancements in training methodologies.